1. Thomas G. Dietterich Department of Computer Science Oregon State University Corvallis, Oregon 97331 http://www.cs.orst.edu/~tgd Ensembles for Cost-Sensitive Learning

2. Outline Cost-Sensitive Learning Problem Statement; Main Approaches Preliminaries Standard Form for Cost Matrices Evaluating CSL Methods Costs known at learning time Costs unknown at learning time Open Problems

3. Cost-Sensitive Learning Learning to minimize the expected cost of misclassifications Most classification learning algorithms attempt to minimize the expected number of misclassification errors In many applications, different kinds of classification errors have different costs, so we need cost-sensitive methods

4. Examples of Applications with Unequal Misclassification Costs Medical Diagnosis: Cost of false positive error: Unnecessary treatment; unnecessary worry Cost of false negative error: Postponed treatment or failure to treat; death or injury Fraud Detection: False positive: resources wasted investigating non- fraud False negative: failure to detect fraud could be very expensive

5. Related Problems Imbalanced classes: Often the most expensive class (e.g., cancerous cells) is rarer and more expensive than the less expensive class Need statistical tests for comparing expected costs of different classifiers and learning algorithms

6. Example Misclassification Costs Diagnosis of Appendicitis Cost Matrix: C(i,j) = cost of predicting class i when the true class is j Predicted State of Patient True State of Patient PositiveNegative Positive 11 Negative 1000

7. Estimating Expected Misclassification Cost Let M be the confusion matrix for a classifier: M(i,j) is the number of test examples that are predicted to be in class i when their true class is j Predicted Class True Class PositiveNegative Positive 4016 Negative 836

8. Estimating Expected Misclassification Cost (2) The expected misclassification cost is the Hadamard product of M and C divided by the number of test examples N: i,j M(i,j) * C(i,j) / N We can also write the probabilistic confusion matrix: P(i,j) = M(i,j) / N. The expected cost is then P * C

9. Interlude: Normal Form for Cost Matrices Any cost matrix C can be transformed to an equivalent matrix C with zeroes along the diagonal Let L(h,C) be the expected loss of classifier h measured on loss matrix C. Defn: Let h 1 and h 2 be two classifiers. C and C are equivalent if L(h 1,C) > L(h 2,C) iff L(h 1,C) > L(h 2,C)

10. Theorem (Margineantu, 2001) Let be a matrix of the form If C 2 = C 1 +, then C 1 is equivalent to C 2 1 2 … k 1 2 … k … 1 2 … k

11. Proof Let P 1 (i,k) be the probabilistic confusion matrix of classifier h 1, and P 2 (i,k) be the probabilistic confusion matrix of classifier h 2 L(h 1,C) = P 1 * C L(h 2,C) = P 2 * C L(h 1,C) – L(h 2,C) = [P 1 – P 2 ] * C

12. Proof (2) Similarly, L(h 1,C) – L(h 2, C) = [P 1 – P 2 ] * C = [P 1 – P 2 ] * [C + ] = [P 1 – P 2 ] * C + [P 1 – P 2 ] * We now show that [P 1 – P 2 ] * = 0, from which we can conclude that L(h 1,C) – L(h 2,C) = L(h 1,C) – L(h 2,C) and hence, C is equivalent to C.

13. Proof (3) [P 1 – P 2 ] * = i k [P 1 (i,k) – P 2 (i,k)] * (i,k) = i k [P 1 (i,k) – P 2 (i,k)] * k = k k i [P 1 (i,k) – P 2 (i,k)] = k k i [P 1 (i|k) P(k) – P 2 (i|k) P(k)] = k k P(k) i [P 1 (i|k) – P 2 (i|k)] = k k P(k) [1 – 1] = 0

14. Proof (4) Therefore, L(h 1,C) – L(h 2,C) = L(h 1,C) – L(h 2,C). Hence, if we set k = –C(k,k), then C will have zeroes on the diagonal

15. End of Interlude From now on, we will assume that C(i,i) = 0

16. Interlude 2: Evaluating Cost- Sensitive Learning Algorithms Evaluation for a particular C: BC OST and BD ELTA C OST procedures Evaluation for a range of possible Cs: AUC: Area under the ROC curve Average cost given some distribution D(C) over cost matrices

17. Two Statistical Questions Given a classifier h, how can we estimate its expected misclassification cost? Given two classifiers h 1 and h 2, how can we determine whether their misclassification costs are significantly different?

18. Estimating Misclassification Cost: BC OST Simple Bootstrap Confidence Interval Draw 1000 bootstrap replicates of the test data Compute confusion matrix M b, for each replicate Compute expected cost c b = M b * C Sort c b s, form confidence interval from the middle 950 points (i.e., from c (26) to c (975) ).

19. Comparing Misclassification Costs: BD ELTA C OST Construct 1000 bootstrap replicates of the test set For each replicate b, compute the combined confusion matrix M b (i,j,k) = # of examples classified as i by h 1, as j by h 2, whose true class is k. Define (i,j,k) = C(i,k) – C(j,k) to be the difference in cost when h 1 predicts class i, h 2 predicts j, and the true class is k. Compute b = M b * Sort the b s and form a confidence interval [ (26), (975) ] If this interval excludes 0, conclude that h 1 and h 2 have different expected costs

20. ROC Curves Most learning algorithms and classifiers can tune the decision boundary Probability threshold: P(y=1|x) > Classification threshold: f(x) > Input example weights Ratio of C(0,1)/C(1,0) for C-dependent algorithms

21. ROC Curve For each setting of such parameters, given a validation set, we can compute the false positive rate: fpr = FP/(# negative examples) and the true positive rate tpr = TP/(# positive examples) and plot a point (tpr, fpr) This sweeps out a curve: The ROC curve

22. Example ROC Curve

23. AUC: The area under the ROC curve AUC = Probability that two randomly chosen points x 1 and x 2 will be correctly ranked: P(y=1|x 1 ) versus P(y=1|x 2 ) Measures correct ranking (e.g., ranking all positive examples above all negative examples) Does not require correct estimates of P(y=1|x)

24. Direct Computation of AUC (Hand & Till, 2001) Direct computation: Let f(x i ) be a scoring function Sort the test examples according to f Let r(x i ) be the rank of x i in this sorted order Let S 1 = {i: yi=1} r(x i ) be the sum of ranks of the positive examples AUC = [S 1 – n 1 (n 1 +1)/2] / [n 0 n 1 ] where n 0 = # negatives, n 1 = # positives

25. Using the ROC Curve Given a cost matrix C, we must choose a value for that minimizes the expected cost When we build the ROC curve, we can store with each (tpr, fpr) pair Given C, we evaluate the expected cost according to 0 * fpr * C(1,0) + 1 * (1 – tpr) * C(0,1) where 0 = probability of class 0, 1 = probability of class 1 Find best (tpr, fpr) pair and use corresponding threshold

26. End of Interlude 2 Hand and Till show how to generalize the ROC curve to problems with multiple classes They also provide a confidence interval for AUC

27. Outline Cost-Sensitive Learning Problem Statement; Main Approaches Preliminaries Standard Form for Cost Matrices Evaluating CSL Methods Costs known at learning time Costs unknown at learning time Variations and Open Problems

28. Two Learning Problems Problem 1: C known at learning time Problem 2: C not known at learning time (only becomes available at classification time) Learned classifier should work well for a wide range of Cs

29. Learning with known C Goal: Given a set of training examples {(x i, y i )} and a cost matrix C, Find a classifier h that minimizes the expected misclassification cost on new data points (x*,y*)

30. Two Strategies Modify the inputs to the learning algorithm to reflect C Incorporate C into the learning algorithm

31. Strategy 1: Modifying the Inputs If there are only 2 classes and the cost of a false positive error is times larger than the cost of a false negative error, then we can put a weight of on each negative training example = C(1,0) / C(0,1) Then apply the learning algorithm as before

32. Some algorithms are insensitive to instance weights Decision tree splitting criteria are fairly insensitive (Holte, 2000)

33. Setting By Class Frequency Set / 1/n k, where n k is the number of training examples belonging to class k This equalizes the effective class frequencies Less frequent classes tend to have higher misclassification cost

34. Setting by Cross-validation Better results are obtained by using cross-validation to set to minimize the expected error on the validation set The resulting is usually more extreme than C(1,0)/C(0,1) Margineantu applied Powells method to optimize k for multi-class problems

35. Comparison Study Grey: CV wins; Black: ClassFreq wins; White: tie 800 trials (8 cost models * 10 cost matrices * 10 splits)

36. Conclusions from Experiment Setting according to class frequency is cheaper gives the same results as setting by cross validation Possibly an artifact of our cost matrix generators

37. Strategy 2: Modifying the Algorithm Cost-Sensitive Boosting C can be incorporated directly into the error criterion when training neural networks (Kukar & Kononenko, 1998)

38. Cost-Sensitive Boosting (Ting, 2000) Adaboost (confidence weighted) Initialize w i = 1/N Repeat Fit h t to weighted training data Compute t = i y i h t (x i ) w i Set t = ½ * ln (1 + t )/(1 – t ) w i := w i * exp(– t y i h t (x i ))/Z t Classify using sign( t t h t (x))

39. Three Variations Training examples of the form (x i, y i, c i ), where c i is the cost of misclassifying x i AdaCost (Fan et al., 1998) w i := w i * exp(– t y i h t (x i ) i )/Z t i = ½ * (1 + c i ) if error = ½ * (1 – c i ) otherwise CSB2 (Ting, 2000) w i := i w i * exp(– t y i h t (x i ))/Z t i = c i if error = 1 otherwise SSTBoost (Merler et al., 2002) w i := w i * exp(– t y i h t (x i ) i )/Z t i = c i if error i = 2 – c i otherwise c i = w for positive examples; 1 – w for negative examples

40. Additional Changes Initialize the weights by scaling the costs c i w i = c i / j c j Classify using confidence weighting Let F(x) = t t h t (x) be the result of boosting Define G(x,k) = F(x) if k = 1 and –F(x) if k = –1 predicted y = argmin i k G(x,k) C(i,k)

41. Experimental Results: (14 data sets; 3 cost ratios; Ting, 2000)

42. Open Question CSB2, AdaCost, and SSTBoost were developed by making ad hoc changes to AdaBoost Opportunity: Derive a cost-sensitive boosting algorithm using the ideas from LogitBoost (Friedman, Hastie, Tibshirani, 1998) or Gradient Boosting (Friedman, 2000) Friedmans MART includes the ability to specify C (but I dont know how it works)

43. Outline Cost-Sensitive Learning Problem Statement; Main Approaches Preliminaries Standard Form for Cost Matrices Evaluating CSL Methods Costs known at learning time Costs unknown at learning time Variations and Open Problems

44. Learning with Unknown C Goal: Construct a classifier h(x,C) that can accept the cost function at run time and minimize the expected cost of misclassification errors wrt C Approaches: Learning to estimate P(y|x) Learn a ranking function such that f(x 1 ) > f(x 2 ) implies P(y=1|x 1 ) > P(y=1|x 2 )

45. Learning Probability Estimators Train h(x) to estimate P(y=1|x) Given C, we can then apply the decision rule: y = argmin i k P(y=k|x) C(i,k)

46. Good Class Probabilities from Decision Trees Probability Estimation Trees Bagged Probability Estimation Trees Lazy Option Trees Bagged Lazy Option Trees

47. Causes of Poor Decision Tree Probability Estimates Estimates in leaves are based on a small number of examples (nearly pure) Need to sub-divide pure regions to get more accurate probabilities

48. Probability Estimates are Extreme Single decision tree; 700 examples

49. Need to Subdivide Pure Leaves P(y=1|x) x 0.5 Consider a region of the feature space X. Suppose P(y=1|x) looks like this:

50. Probability Estimation versus Decision-making P(y=1|x) x 0.5 predict class 0 predict class 1 A simple CLASSIFIER will introduce one split

51. Probability Estimation versus Decision-making P(y=1|x) x 0.5 A PROBABILITY ESTIMATOR will introduce multiple splits, even though the decisions would be the same

52. Probability Estimation Trees (Provost & Domingos, in press) C4.5 Prevent extreme probabilities: Laplace Correction in the leaves P(y=k|x) = (n k + 1/K) / (n + 1) Need to subdivide: no pruning no collapsing

53. Bagged PETs Bagging helps solve the second problem Let {h 1, …, h B } be the bag of PETs such that h b (x) = P(y=1|x) estimate P(y=1|x) = 1/B * b h b (x)

54. ROC: Single tree versus 100- fold bagging

55. AUC for 25 Irvine Data Sets (Provost & Domingos, in press)

56. Notes Bagging consistently gives a huge improvement in the AUC The other factors are important if bagging is NOT used: No pruning/collapsing Laplace-corrected estimates

57. Lazy Trees Learning is delayed until the query point x* is observed An ad hoc decision tree (actually a rule) is constructed just to classify x*

58. Growing a Lazy Tree (Friedman, Kohavi, Yun, 1985) x 1 > 3 x 4 > -2 Only grow the branches corresponding to x* Choose splits to make these branches pure

59. Option Trees (Buntine, 1985; Kohavi & Kunz, 1997) Expand the Q best candidate splits at each node Evaluate by voting these alternatives

60. Lazy Option Trees (Margineantu & Dietterich, 2001) Combine Lazy Decision Trees with Option Trees Avoid duplicate paths (by disallowing split on u as child of option v if there is already a split v as a child of u): vu v u

61. Bagged Lazy Option Trees (B-LOTs) Combine Lazy Option Trees with Bagging (expensive!)

62. Comparison of B-PETs and B-LOTs Overlapping Gaussians Varying amount of training data and minimum number of examples in each leaf (no other pruning)

63. B-PET vs B-LOT Bagged PETs give better ranking Bagged LOTs give better calibrated probabilities Bagged PETsBagged LOTs

64. B-PETs vs B-LOTs Grey: B-LOTs win Black: B-PETs win White: Tie Test favors well-calibrated probabilities

65. Open Problem: Calibrating Probabilities Can we find a way to map the outputs of B-PETs into well-calibrated probabilities? Post-process via logistic regression? Histogram calibration is crude but effective (Zadrozny & Elkan, 2001)

66. Comparison of Instance-Weighting and Probability Estimation Black: B-PETs win; Grey: ClassFreq wins; White: Tie

67. An Alternative: Ensemble Decision Making Dont estimate probabilities: compute decision thresholds and have ensemble vote! Let = C(0,1) / [C(0,1) + C(1,0)] Classify as class 0 if P(y=0|x) > Compute ensemble h 1, …, h B of probability estimators Take majority vote of h b (x) >

68. Results (Margineantu, 2002) On KDD-Cup 1998 data (Donations), in 100 trials, a random-forest ensemble beats B-PETs 20% of the time, ties 75%, and loses 5% On Irvine data sets, a bagged ensemble beats B-PETs 43.2% of the time, ties 48.6%, and loses 8.2% (averaged over 9 data sets, 4 cost models)

69. Conclusions Weighting inputs by class frequency works surprisingly well B-PETs would work better if they were well-calibrated Ensemble decision making is promising

70. Outline Cost-Sensitive Learning Problem Statement; Main Approaches Preliminaries Standard Form for Cost Matrices Evaluating CSL Methods Costs known at learning time Costs unknown at learning time Open Problems and Summary

71. Open Problems Random forests for probability estimation? Combine example weighting with ensemble methods? Example weighting for CART (Gini) Calibration of probability estimates? Incorporation into more complex decision- making procedures, e.g. Viterbi algorithm?

72. Summary Cost-sensitive learning is important in many applications How can we extend discriminative machine learning methods for cost-sensitive learning? Example weighting: ClassFreq Probability estimation: Bagged LOTs Ranking: Bagged PETs Ensemble Decision-making

73. Bibliography Buntine, W. 1990. A theory of learning classification rules. Doctoral Dissertation. University of Technology, Sydney, Australia. Drummond, C., Holte, R. 2000. Exploiting the Cost (In)sensitivity of Decision Tree Splitting Criteria. ICML 2000. San Francisco: Morgan Kaufmann. Friedman, J. H. 1999. Greedy Function Approximation: A Gradient Boosting Machine. IMS 1999 Reitz Lecture. Tech Report, Department of Statistics, Stanford University. Friedman, J. H., Hastie, T., Tibshirani, R. 1998. Additive Logistic Regression: A Statistical View of Boosting. Department of Statistics, Stanford University. Friedman, J., Kohavi, R., Yun, Y. 1996. Lazy decision trees. Proceedings of the Thirteenth National Conference on Artificial Intelligence. (pp. 717-724). Cambridge, MA: AAAI Press/MIT Press.

74. Bibliography (2) Hand, D., and Till, R. 2001. A Simple Generalisation of the Area Under the ROC Curve for Multiple Class Classification Problems. Machine Learning, 45(2): 171. Kohavi, R., Kunz, C. 1997. Option decision trees with majority votes. ICML-97. (pp 161-169). San Francisco, CA: Morgan Kaufmann. Kukar, M. and Kononenko, I. 1998. Cost-sensitive learning with neural networks. Proceedings of the European Conference on Machine Learning. Chichester, NY: Wiley. Margineantu, D. 1999. Building Ensembles of Classifiers for Loss Minimization, Proceedings of the 31st Symposium on the Interface: Models, Prediction, and Computing. Margineantu, D. 2001. Methods for Cost-Sensitive Learning. Doctoral Dissertation, Oregon State University.

75. Bibliography (3) Margineantu, D. 2002. Class probability estimation and cost-sensitive classification decisions. Proceedings of the European Conference on Machine Learning. Margineantu, D. and Dietterich, T. 2000. Bootstrap Methods for the Cost- Sensitive Evaluation of Classifiers. ICML 2000. (pp. 582-590). San Francisco: Morgan Kaufmann. Margineantu, D., Dietterich, T. G. 2002. Improved class probability estimates from decision tree models. To appear in Lecture Notes in Statistics. New York, NY: Springer Verlag. Provost, F., Domingos, P. In Press. Tree induction for probability-based ranking. To appear in Machine Learning. Available from Provost's home page. Ting, K. 2000. A comparative study of cost-sensitive boosting algorithms. ICML 2000. (pp 983-990) San Francisco, Morgan Kaufmann. (Longer version available from his home page.)

76. Bibliography (4) Zadrozny, B., Elkan, C. 2001. Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers. ICML-2001. (pp 609- 616). San Francisco, CA: Morgan Kaufmann.